Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

tensorflow: better handling name_scope #16

Merged
merged 2 commits into from
Jun 28, 2019
Merged

tensorflow: better handling name_scope #16

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
14a4377
tensorflow: better handling name_scope
bobzhuyb Jun 28, 2019
2ad03b9
Merge branch 'master' into tf_scope_fix
ymjiang Jun 28, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 6 additions & 4 deletions byteps/tensorflow/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,23 +56,25 @@ def push_pull(tensor, scope='', average=True, device_dense='', device_sparse='',
return new_tensor


def broadcast_global_variables(root_rank):
def broadcast_global_variables(root_rank, scope=''):
"""Broadcasts all global variables from root rank to all other processes.
Arguments:
root_rank: rank of the process from which global variables will be broadcasted
to all other processes.
scope: the graph name scope
"""
return broadcast_variables(tf.global_variables(), root_rank)
return broadcast_variables(tf.global_variables(), root_rank, scope)


def broadcast_variables(variables, root_rank):
def broadcast_variables(variables, root_rank, scope=''):
"""Broadcasts variables from root rank to all other processes.
Arguments:
variables: variables for broadcast
root_rank: rank of the process from which global variables will be broadcasted
to all other processes.
scope: the graph name scope
"""
return tf.group(*[tf.assign(var, broadcast(var, root_rank))
return tf.group(*[tf.assign(var, broadcast(var, root_rank, scope))
for var in variables])


Expand Down
10 changes: 8 additions & 2 deletions byteps/tensorflow/ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,6 +79,8 @@ def _push_pull(tensor, scope='', name=None):
"""
if name is None and not _executing_eagerly():
name = 'BytePSPushPull_%s' % _normalize_name(tensor.name)
if scope == '' and not _executing_eagerly():
scope = tf.compat.v1.get_default_graph().get_name_scope()
full_name = scope + name
full_name = full_name.encode("ascii")
TF_LIB_CTYPES.byteps_tensorflow_declare_tensor(ctypes.c_char_p(full_name))
Expand All @@ -97,7 +99,7 @@ def _push_pull_grad(op, grad):
return _push_pull(grad)


def broadcast(tensor, root_rank, name=None, is_variable=True):
def broadcast(tensor, root_rank, scope='', name=None, is_variable=True):
"""An op which broadcasts the input tensor on root rank to the same input tensor
on all other BytePS processes.
The broadcast operation is keyed by the name of the op. The tensor type and
Expand All @@ -110,7 +112,11 @@ def broadcast(tensor, root_rank, name=None, is_variable=True):
# Broadcast is implemented as push + pull after zero-ing non-root tensors
if name is None and not _executing_eagerly():
name = 'BytePSBroadcast_%s' % _normalize_name(tensor.name)
TF_LIB_CTYPES.byteps_tensorflow_declare_tensor(ctypes.c_char_p(name.encode("ascii")))
if scope == '' and not _executing_eagerly():
scope = tf.compat.v1.get_default_graph().get_name_scope()
full_name = scope + name
full_name = full_name.encode("ascii")
TF_LIB_CTYPES.byteps_tensorflow_declare_tensor(ctypes.c_char_p(full_name))
if is_variable and (root_rank != rank()):
return C_LIB.byteps_push_pull(tensor.assign(tf.zeros_like(tensor)), name=name)
else:
Expand Down